fundamental*uncertain/es*in*the*science* · confidence,*uncertainty*and*decision3support...
TRANSCRIPT
Fundamental Uncertain/es in the Science
Brad Udall -‐ [email protected] Western Water Assessment CIRES, University of Colorado
Planning for Local Government Climate Challenges: Connec/ng Research and Prac/ce Tempe, AZ October 18, 2012
Model
Climate Variability
Scenario
Model
Scenario
Climate Variability
Bri/sh Isles Global
Hawkins and SuUon, 2009
Talk Overview • Concepts re Uncertainty
– On Policy Making -‐ Smith and Stern, 2011 – On Climate Modeling -‐ Stainforth et al., 2007 – Known Issues – GCMs, Downscaling, Natural Variability, etc – Warnings from Nature and Science – Type 1 vs. Type 2 Errors
• Case Studies on Uncertainty, Projec/ons – Australia via Kiem & Verdon-‐Kidd – Uncertainty – Morgan and Henrion – USGCRP SAP 5.2 on Uncertainty – NRC Report Informing Decisions
• Conclusions and Musings • Note: Focus is on Climate Change and Modeling, but these concepts apply to all
uncertainty issues and modeling • Advice From Churchill…Doing the Right Thing…
Uncertainty in Science and its role in Climate Policy
• Terrific, Thought-‐Provoking Ar/cle Worth a Read • Key Points
– Science focuses on what is known or almost known • What is unlikely to be known can also aid policy maker
– Large Uncertain/es do not mean small risks • Uncertainty can support immediate ac/on in some cases
– A lack of certainty provides no ra/onal argument against ac/on – Varie/es of uncertainty:
• Imprecision – can be quan/fied by PDF • Ambiguity – impacts known but can’t be quan/fied via PDF, e.g. 100 yr impacts • Intractability – not solvable, e.g. no equa/ons or lack computers • Indeterminacy – also not solvable, e.g., a societal value or non-‐physical parmeter
– “Models can increase our understanding long before they start providing realis/c numbers.”
– Nice Discussion about Hydrological Uncertain/es – These concepts not appreciated by both modeling community and Decision-‐
Making community
Uncertainty in science and its role in climate policy Leonard Smith and Nicholas Stern Phil. Trans. R. Soc. A (2011) 369, 1–24 doi:10.1098/rsta.2011.0149
Confidence, Uncertainty and Decision-‐Support Relevance in Climate Predic/ons
• Models can’t be calibrated – simula/ng never before seen state – Contrast with Weather where models interpolate – Climate models: no archive over /me, run once, projec/on /mes >> model life
• PDFs can be made but what they represent is not what we thought – We thought a ‘PDF of the Future’ ini/ally – But models are only a Lower Bound on Range of Uncertainty – True Future PDF is wider
• At least 3 Sources of Uncertainty – Forcing
• Emissions Scenario – Ini/al Condi/ons
• Makes a Difference to End Results • Does Not Make a Difference to End Results
– Model Imperfec/on • Model Uncertainty – e.g. parameters for physical processes • Model Inadequacy – incorrect formula/on
• No ra/onal way to weight models now
Confidence, uncertainty and decision-‐support relevance in climate predic;ons D.A Stainforth, M.R Allen, E.R Tredger and L.A Smith Phil. Trans. R. Soc. A 2007 365, 2145-‐2161 doi: 10.1098/rsta.2007.2074
Some Known GCM -‐ LSM Uncertain/es
• Emissions Scenarios • Models show consistent errors (biases) • Known Problems: El Nino, Monsoons,
Mountains, Blocking, Extremes • Natural Variability Cri/cal – inherent
limits to predictability exist • Models Constantly Changing • Can’t Verify Model Output • Won’t get any beUer for at least 10
years – See WUCA -‐-‐ and maybe not even then…
• How much hydro cycle enhanced? • Downscaling Issues
– CRB Flows can change by > 10% • Land Surface Models
– Sensi/vity to T & P Varies “The First Principle is you must not fool yourself and you are the easiest person to fool.” ~ Feynman
Regional Climate Change Assessments
Recent Science Ar3cles on Similar Themes
“Many regional modelers don’t do an adequate job of quan/fying issues of uncertainty.” “We are not confident predic/ng the things people are most interested in being predicted.” “The problem is that precision is oren mistaken for accuracy.” ~ Christopher Bretherton University of Washington
The 4 Outcomes and 2 Types of Errors from Predic/ons
Note: Which error is worse is en/rely contextual
Diagram shows Predic/ons versus actual outcomes. ScaUer represents uncertainty. Step 1: Pick a threshold for something bad to occur, e.g., Wind Speeds during Hurricane
Step 2: Pick a Decision Cutoff for Societal Ac/on Here: Cutoff same as threshold at 50
You’ve now determined the four outcomes: True Posi/ve, True Nega/ve, Cry Wolf, Trojan Horse
Lessons from Taylor Russell Diagrams
Cry Wolf
Trojan Horse
Lessons From Taylor – Russell Diagrams on Errors
Cry Wolf
Trojan Horse
Trojan Horse
Cry Wolf
By Adjus/ng the Decision Cutoff you can reduce either Cry Wolf Errors or Trojan Horse Errors but not both at the same /me. Top: Decision Cutoff at 65 Minimizes Cry Wolf Errors at the expense of increasing Trojan Horse Errors
BoUom: Decision Cutoff at 35 Minimizes Trojan Horse Errors at the expense of increasing Cry Wolf Errors
Lessons from Taylor-‐Russell Diagrams: Can’t Reduce Cry Wolf and Trojan Horse at the Same Time
In response to poorly perceived predic/on outcomes, society some/mes oscillates over /me between different Decision Cutoffs and hence rate of the two errors.
Talk Overview • Concepts re Uncertainty
– On Policy Making -‐ Smith and Stern, 2011 – On Climate Modeling -‐ Stainforth et al., 2007 – Known Issues – GCMs, Downscaling, Natural Variability, etc – Warnings from Nature and Science – Type 1 vs. Type 2 Errors
• Case Studies on Uncertainty, Projec/ons – Australia via Kiem & Verdon-‐Kidd – Uncertainty – Morgan and Henrion – USGCRP SAP 5.2 on Uncertainty – NRC Report Informing Decisions
• Conclusions and Musings • Note: Focus is on Climate Change and Modeling, but these concepts apply to all
uncertainty issues and modeling • Advice From Churchill…Doing the Right Thing…
An Australian Example – Kiem and Verdon-‐Kidd
KVK -‐ Iden/fied Shortcomings with Current Approach
• Failure to simulate synop/c paUerns that drive rainfall, especially extremes
• Large Scale Processes not well simulated: ENSO, IOD, Others – Not understood, either
• Of 39 GCM runs, 22 show increases, 17 show decreases in precipita/on • None of the models could reproduce the drying trend since mid 1990s • GCMs couldn’t dis/nguish between wet coastal strip and dry interior 300
km away • Climate model outputs at monthly and submonthly scale do not
reproduce historical climate and show significant biases – Downscaling (Bias Correc/on and Change Factor) introduce ‘false precision’ and
introduce an addi/onal layer of uncertainty – Bias Correc/ons assumed to be sta/onary over /me
• Climate to Hydrology Connec/on is Poorly Understood – Current Hydrology models calibrated to current condi/ons, not future
• When does such calibra/on cease to be useful?
KVK -‐ “A 5-‐Step Way Forward” • Step 1: Communica/on Between Climate Scien/sts, Hydrologists, and Water
Resource Managers – Define what is ‘prac/cally useful’ – Disconnects about what can be expected and How to act in face of uncertainty
• Step 2: Quan/fy Baseline Risk Associated with Natural Climate Variability – Need to understand paleoclimate beUer
• How dry can it get and for how long?
– Need to understand drivers of variability – Stochas/c Framework needed to integrate both
• Step 3: Incorporate the Projected Impacts of Anthropogenic Change – Iden/fy physical processes driving hydroclimate – Iden/fy or Develop models that simulate these processes – Determine how processes will change in the future and apply changes in stochas/c
framework
• Step 4: Develop Appropriate Adapta/on Strategies – Need Reliable Probabili/es of Uncertain/es – Robust Quan/fica/on of Uncertain/es Needed – Iden/fy Win-‐win Adapta/on Strategies
• Step 5: Ongoing Communica/on – Address Intensity, Time Span, Frequency
1. Do your homework with literature, experts and users • Iterate, reframing as necessary
2. Let the problem drive the analysis • Avoid hammers looking for nails
3. Make the analysis as simple as possible but not simpler • Simpler: can explain easier • No Simpler: don’t overlook important detail
4. Iden/fy all significant assump/ons • Policy concerns and decisions
5. Be explicit about decision criteria and policy strategies • Subset of 4 above, but cri/cal to note
6. Be explicit about uncertain/es • Technical, Scien/fic, Economic, Poli/cal • Model Func/onal Form • Disagreements among Experts
7. Perform Systema/c sensi/vity and uncertainty analysis • Vary inputs in models to expose sensi/vi/es
8. Itera/vely refine the problem statement and the analysis • Add and/or Subtract from analysis if sensi/vity so indicates
9. Document clearly and completely • Too important to leave un/l the end
10. Expose the work to peer review • Yes, even policy analysis, not just technical work
10 Commandments for Good Policy Analysis
Morgan and Henrion, 1990
Lessons from SAP 5.2 – Uncertainty Guidance for Researchers
• Does what we are doing make sense?
• Are there other important factors that are equally or more important than the factors we are considering?
• Are there key correla/on structures in the problem that are being ignored?
• Are there standard assump/ons and judgments about which we are not being explicit?
• Is informa/on about the uncertain/es related to research results and poten/al policies being communicated clearly and consistently?
M. Granger Morgan, Hadi Dowlatabadi, Max Henrion, David Keith, Robert Lempert, Sandra McBride, Mitchell Small, and Thomas Wilbanks
Informing Decisions in a Changing Climate, NRC 2009 • EPA and NOAA asked NRC to study informing decisions
• Sta/onary Climate Assump/on No Longer Tenable
• 6 Key Principles • Begin with Users Needs • Give Priority to Process over Products • Link Informa/on Users and Providers • Build Connec/ons Across Disciplines • Seek Ins/tu/onal Support • Design Process for Learning
• Other Ideas • Use “Delibera/on with Analysis” • Broadly Scan for New Informa/on, Including Overseas • USGCRP Should support Research on Decisions, not just climate • Maintain and Expand Observing Systems • Support Students with Specialized Decision Support Knowledge • Undertake Na/onal Ini/a/ve for Decision Support
• Note: Almost nothing on forecasts, projec/ons
Some Musings and Conclusions -‐ 1 • Churchill’s Advice – Need to Think! • Water Users have goUen smarter, not so sure about some researchers
– WUCA a case in point – No one, far as I know, is using hydrologic outputs to make decisions – 1 km downscaling by ecological researchers
• Science Needs to Have BeUer, Much BeUer Communica/on with Decision Makers – Push Back on What we Can and Cannot Do with Decision Makers – We must show full range of futures and explain different risks with each – credibility! – Uncertainty can in fact be a cause for ac/on
• Need to Have BeUer Language around Uncertainty – Not all uncertainty the same – Like Smith & Stern Language: Imprecision, Ambiguity, Intractable, Irreducible – Need to discuss Emissions, Variability, El Nino etc in these terms in addi/on to other
knowledge about these effects – Can we create a food label for these projects showing uncertainty? – Avoid being overly determinis/c and precise when there is no basis other than our
ability to produce model output of unknown quality?
Some Musings and Conclusions -‐ 2 • Projec/ons least valuable part of modeling
– “We Model for Insights not Answers” – “Plans are Nothing, Planning is Everything” – Physical Mechanisms, Counter-‐intui/ve findings, Data Issues, Learning ALL trump projec/ons
• Modeling as thinking… – Wri/ng is thinking with words, modeling is thinking with numbers – Wri/ng generates mul/ple insights, so should modeling – Turning the crank to blindly produce outputs is not thinking – Simple Models are ok, too
• Cry Wolf versus Trojan Horse Errors – Be aware that lowering one raises the other – Think about which kind of error is more important to avoid for the issue at hand – Science historically has focused on avoiding Cry Wolf
• Specializa/on leads to blindness – Are Really Smart People doing collec/vely dumb things? – We need to bring other disciplines into this discussion – The less you want to read Smith and Stern, Stainforth et al, the more you need to read them!
• Easy to Dump on Models, not so easy to define a healthy rela/onship – This is our challenge
A WeUer Colorado River?
q Unforced internal model variability and model choice dominate empirical CDFs of streamflow change projec/ons
Note: not much difference by future periods
At 3 Future Periods At 2070-‐2099
Note: limited difference by emissions scenarios
Slide from Andy Wood
Rasmussen et al., 2011: increased Precip due to beUer topography Harding et al., 2012: Don’t forget 30% of runs which show greater precip
Mysteries vs. Puzzles Two Kinds of Problems Malcolm Gladwell on Enron Collapse
Mystery: problem where the answer is obscured by lots of details. The answer is there but most of us can’t see it. Puzzle: problem where we are missing a crucial piece of information. We need to search for the information. The solutions required are very different. View climate as a mystery, not as a puzzle.
Black Swans
• Society Changing Event – Outlier and Extreme – 2008 Market Crash – Gulf Oil Spill – 9/11 – 2005 Tsunami
• Happen all the time and change history • We fool ourselves into thinking that we can predict these • Instead of predicting, need to adjust to existence • We fixate on what we know, rather than what we don’t know
– Knowledge makes us more confident than we should we
• Will occur more frequently in 21st century • Taleb loathes normal distribution
– Poorly describes risk profile for many vulnerabilities – ‘Fat Tails’ Exist in many phenomenon
• Name comes from Australia and first sightings of Black Swans
If you increase accuracy of predic/ons you can reduce Both Cry Wolf and Trojan Horse Errors at the same /me Oren, however, this is not an op/on.
KVK -‐ Key Quote re Projec/ons of Extremes
• “Un/l climate models can be shown to sa/sfactorily simulate the physical mechanisms we know are important for regional scale hydrology, their outputs should be used with cau/on as there is the danger that the “worst-‐case scenario” projected by current climate models may not actually be the worst case possible for water resource managers in terms of extreme events, par/cularly if the baseline is inappropriately assessed [e.g., Verdon-‐Kidd and Kiem, 2010].
• This has been recently demonstrated in eastern Australia with the December 2010 and January 2011 flooding and prior to that with the Big Dry drought, which lasted more than a decade; both of these events were far more severe than anything projected by any climate models under even the worst emissions scenarios and out to at least 2070.”
Deser et al, Uncertainty in Backyard • “While there is undoubtedly much room for model
improvement, we show that natural climate variability poses inherent limits to climate predictability and the related goal of adapta/on guidance at many places in North America.”
• “On the other hand, low natural variability in some loca/ons leads to a more predictable future in which anthropogenic forcing can be much more readily iden/fied, even on small scales.”
• “We call for a more focused dialogue between scien/sts, policymakers, and the public to improve communica/on, avoid raising expecta/ons for accurate regional predic/ons everywhere, and recognize that in some loca/ons more useful predictability can be expected.”
• Harding et al on CRB Projec/ons: at Mid-‐century one GCM with 4 runs shows -‐30% and +30%
“Uncertainty in the Backyard: Communica;ng the Role of Natural Variability in Future North American Climate” Clara Deser, Reto Knu}, Susan Solomon and Adam S. Phillips January 17, 2012, Perspec3ve submi@ed to Nature Climate Change